Hindawi

Mathematical Problems in Engineering

Volume 2017, Article ID 8061561, 11 pages
https://doi.org/10.1155/2017/8061561

Research Article
An Efficient Implementation of Track-Oriented Multiple
Hypothesis Tracker Using Graphical Model Approaches

Jinping Sun,1 Qing Li,1 Xuwang Zhang,1 and Wei Sun2

1
School of Electronic and Information Engineering, Beihang University, Beijing 100191, China
2
Nanjing Electronic Technology Research Institute, Nanjing 210039, China

Correspondence should be addressed to Jinping Sun; [email protected]

Received 27 March 2017; Accepted 13 August 2017; Published 11 September 2017

Academic Editor: Federica Caselli

Copyright © 2017 Jinping Sun et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

The multiple hypothesis tracker (MHT) is currently the preferred method for addressing data association problem in multitarget
tracking (MTT) application. MHT seeks the most likely global hypothesis by enumerating all possible associations over time, which
is equal to calculating maximum a posteriori (MAP) estimate over the report data. Despite being a well-studied method, MHT
remains challenging mostly because of the computational complexity of data association. In this paper, we describe an efficient
method for solving the data association problem using graphical model approaches. The proposed method uses the graph repre-
sentation to model the global hypothesis formation and subsequently applies an efficient message passing algorithm to obtain the
MAP solution. Specifically, the graph representation of data association problem is formulated as a maximum weight independent
set problem (MWISP), which translates the best global hypothesis formation into finding the maximum weight independent set on
the graph. Then, a max-product belief propagation (MPBP) inference algorithm is applied to seek the most likely global hypotheses
with the purpose of avoiding a brute force hypothesis enumeration procedure. The simulation results show that the proposed
MPBP-MHT method can achieve better tracking performance than other algorithms in challenging tracking situations.

1. Introduction there are two types of MHT framework. The first, known as
hypothesis-oriented MHT (HOMHT), is proposed by Reid
Multitarget tracking (MTT) is a crucial component for in [1]. HOMHT directly maintains and propagates global
surveillance systems which aimed to obtain the sequential hypotheses, and the complexity of treating global hypothe-
estimation of the number of targets and their states (positions, ses makes it difficult to be implemented. The alternative
velocities, etc.). The main task of MTT is to partition the framework, which is more favored currently, is track-oriented
received observations into tracks that correspond to correct MHT (TOMHT). Rather than maintaining global hypotheses
targets, and this procedure is also known as the data asso- formed previously, TOMHT reforms global hypotheses using
ciation. The usual constraint is that each target at each scan newly updated tracks on each scan and typically maintains a
gives rise to at most one observation. However, it is not known set of potential tracks using the track tree structure [2].
which observation originates from which target, and there are Even though TOMHT appears to be the most-preferred
as well false observations that are not originated by targets. tracking algorithm, it inherently suffers from the combi-
To address the data association problem, several methods, national explosion that exists in the hypothesis formation
including global nearest neighbor (GNN), joint probabilistic step. To improve the efficiency of generating hypotheses, a
data association (JPDA), and multiple hypothesis tracking multidimensional assignment (MDA) method was proposed
(MHT), were proposed. Among these methods, MHT is and the computational feasibility has been greatly enhanced
widely regarded as the most prominent method for structur- by the use of Lagrangian relaxation [3].
ing the data association problem in MTT systems. MHT is a In addition, graphical models, which have received a lot
deferred decision logic method that allows a firm decision to of attention in recent years, were also introduced to tackle
be postponed until more report data are available. Basically, the data association problem. Graphical models are powerful
2 Mathematical Problems in Engineering

New reports definition of MWISP and also basic concepts of the graphical
models. In particular, we introduce a case of graphical models
Gating Track prediction named Markov Random Field (MRF). Additionally, we also
review an efficient inference algorithm called max-product
belief propagation, which is used in graphical models for cal-
Track formation N-scan pruning
culating MAP assignment. Section 4 outlines the MWISP for-
mulation of data association problem and presents in detail
the MPBP algorithm applied in the proposed MPBP-MHT
Generate Best method. Section 5 shows the empirical results of several algo-
Clustering
global hypothesis rithms about tracking performance over challenging situa-
tions, followed by concluding comments in Section 6.
User

Figure 1: Basic elements of TOMHT algorithm. 2. Track-Oriented Multiple

Hypothesis Tracker
tools to model the joint probability distribution of multiple The proposed method is based on the TOMHT framework,
random variables, in which message passing algorithms can which takes advantage of the track tree structure to manage
be used to solve the inference problem efficiently. Several and maintain hypotheses sets. Figure 1 shows the basic ele-
graph approaches that emerge in the tracking literature were ments of a typical TOMHT system. Hypotheses are reformed
reviewed in [4]. A factor graph aided multiple hypothesis from tracks at each scan and the tracks that survive pruning
tracking approach was proposed to solve the association are predicted to the next scan where the process continues.
probability using the sum-product algorithm [5]. A graph- An overview of the core components of TOMHT framework
theoretic interpretation was introduced in [6] and specif- given in Sections 2.1 and 2.2 provides some efficient strategies
ically, the data association problem was formulated as the for limiting the number of hypotheses.
maximum weight independent set problem (MWISP), which
is a well-studied combinatorial optimization problem and 2.1. Background. In the TOMHT approach, tracks are
known to be NP-hard. To construct suboptimal solutions for updated using all newly received observations within the gate.
the MWISP, a greedy randomized adaptive search procedure Each track is defined by a sequence of observations with the
(GRASP) was developed and applied to MHT framework [7]. restriction that at most one observation is included at each
𝑚
Despite the fact that GRASP based approach is not likely to scan. Let {𝑧𝑖𝑘𝑘 }𝑖𝑘𝑘=1 denote the set of observations received at
find the optimal solution, however, this approach provides a scan 𝑘, where 𝑚𝑘 is the total number of observations and
compact and efficient representation of multiple hypotheses 𝑧𝑖𝑘𝑘 represents the 𝑖𝑘 th observation in scan 𝑘. Then the 𝑗𝑘 th
in MHT framework.
track is denoted as T𝑘𝑗𝑘 = {𝑧𝑖1𝑘 , . . . , 𝑧𝑖𝑘𝑘 }. For instance, in Fig-
In this paper, we utilize an efficient TOMHT implemen-
tation, referred to as MPBP-MHT, to solve the data associa- ure 2, track 2 can be represented as {𝑧1𝑘−2 , 𝑧3𝑘−1 , 𝑧2𝑘 }. Dummy
tion problem by exploiting the graphical models and message observation in scan 𝑘 is denoted as 𝑧0𝑘 with the purpose of
passing algorithms. In our approach, hypothesis generation introducing tracks with missed detection. Tracks are defined
can be represented using a graph structure with nodes repre- as compatible if they do not share common observations; that
senting tracks and edges defining compatibility restrictions, is, one observation can be associated with at most one track.
and in this way finding the best hypothesis is transformed into For the convenience of clustering and solving the MWISP,
seeking the maximum weight independent set on the graph, an incompatibility list (ICL) for each track is maintained,
which is a maximum a posteriori (MAP) problem. Because which contains all existing tracks with which a given track
finding the maximum weighted independent set is known is incompatible.
to be NP-hard for an arbitrary graph and the nature equiv- TOMHT uses a track tree structure to store all possible
alence between MWISP and MAP assignment, as a result, track hypotheses. A track tree represents all possible track
approximate algorithms are introduced to efficiently address hypotheses of the corresponding target. A family of tracks is
this problem. Max-product belief propagation (MPBP), as an composed of multiple tracks corresponding to the same tar-
efficient inference algorithm, is generally used for finding the get, i.e., a set of tracks all emanating from a single root obser-
MAP assignment in a joint probability distribution repre- vation. Figure 2 shows the growth of track trees with scan
sented by a graphical model [8]. Therefore, this paper com- time and the process of global hypotheses generation. When
bines the MWISP formulation and the MPBP inference algo- new scans arrive, the track trees are extended to include new
rithm with the purpose of generating the best hypothesis effi- observations. As the track tree is formed by a set of tracks that
ciently. The effectiveness of the proposed approach is tested share a common root node, all these tracks are incompatible.
over challenging tracking cases and a comparison with A global hypothesis is the subset of all possible tracks that
GRASP-MHT algorithm is made to demonstrate the effi- are compatible. As a result, a global hypothesis can only
ciency of MPBP-MHT approach. include no more than one track from each track tree (see, e.g.,
The rest of this paper is organized as follows: Section 2 Figure 2).
briefly reviews the standard TOMHT framework and several In order to evaluate a certain track hypothesis, a track
techniques to maintain track hypotheses. Section 3 gives the score, which is used to describe the validity of a track, is
Mathematical Problems in Engineering 3

Scan k − 2 Scan k − 1 Scan k

Track 1
z1k−1 z1k Hypothesis 1
Track tree 1
z1k−2 Track 2
z3k−1 z2k

Track tree 2 Track 3

z2k−2 z1k−1 z3k
Hypothesis 2

Track 4
z2k−1 z1k
Track tress 3 ..
z3k−2 .
Track 5
z3k−1 z3k

Figure 2: Formation of hypotheses from tracks in track trees.

associated with each track. A track score is generally defined 2.2. Efficient Mechanisms for Implementation. Due to the
as the log likelihood ratio of the probability of the track being fact that the global hypothesis is made by a combination
generated from true target returns to the probability of all of multiple tracks, the TOMHT may suffer from a heavy
observations being false alarms [9]. A recursive formula for computational burden as there is a potential combination
the track score 𝐿(𝑘) at scan 𝑘 is explosion in the number of hypotheses with the growth of
track trees. As a result, several mechanisms are needed to
𝐿 (𝑘) = 𝐿 (𝑘 − 1) + Δ𝐿 (𝑘) , (1) suppress the number of hypotheses in real-time systems.
One crucial technique for limiting the growth of track
where the track score increment Δ𝐿(𝑘) is
trees is to apply the n-scan pruning strategy. Upon finding
the best global hypothesis, the n-scan pruning strategy prunes
{ln (1 − 𝑃𝐷) if no update on scan 𝑘
Δ𝐿 (𝑘) = { (2) all tracks belonging to a track tree with depth more than 𝑛
Δ𝐿 if track update on scan 𝑘.
{ 𝑢 (𝑘)
and tracks that fail to share a common root observation with
any track included in the best global hypothesis. Another
As the detection probability 𝑃𝐷 is less than unity, it is obvious important strategy to reduce the complexity is clustering,
that ln(1 − 𝑃𝐷) < 0. Therefore, the track score suffers from a which successfully divides a large problem into independent
decrease when the track is not updated. When an observation small problems. A cluster is composed of several track trees
is used to update the track on scan 𝑘, however, the track that share one or more common observations. By clustering,
score increases by the amount Δ𝐿 𝑢 (𝑘). The magnitude of the the global hypothesis formulation step can be realized in
increment Δ𝐿 𝑢 is the sum of kinematic and signal-related individual clusters, which in turn greatly reduce the size of
terms [10]. Assuming the tracker to be a Kalman filter with track hypotheses and computing complexity. Other standard
Gaussian innovations, the explicit form of the increment Δ𝐿 𝑢 , techniques used in TOMHT framework, including gating,
in the case where the only signal-related datum is that a Shiryayev sequential probability ratio test (SSPRT), hypothe-
detection or a miss occurred, is given by sis pruning, and track merging, could be referred to [11, 12].

𝑃𝐷 𝑑2 3. Mathematical Preliminaries
Δ𝐿 𝑢 = ln [ ]− , (3)
(2𝜋)𝑀/2 𝜆 𝑓𝑎 √|S| 2
Graphical models provide a common structure on which
where the time index 𝑘 has been dropped and the following generic inference algorithms can operate. In this section,
notations are used: firstly we give a brief review of MWISP, which will be
used to construct the graphical model in the proposed
𝑀: observation dimension method. Then we give some basic concepts of graphical
𝜆 𝑓𝑎 : false target density models. Particularly, we are interested in a special graphical
model, namely, the Markov Random Field (MRF), as it is
S: observation residual covariance matrix the prototype on which we develop our graphical model for
𝑑2 : normalized statistical distance for the observation tracking applications. Furthermore, we review the standard
defined in terms of observation residual vector ̃z and message passing algorithm for inference problems in graph-
covariance matrix S. ical models. Specifically, we introduce in detail the max-
product belief propagation (MPBP) inference which can be
𝑑2 = ̃z𝑇 S−1 ̃z. (4) used in the loopy graph situations.

Having defined a score for each track, one can determine 3.1. Maximum Weight Independent Set Problem. MWISP is a
the score of a global hypothesis and prunes a track once its well-studied combinatorial optimization problem. Let 𝐺 =
track score falls below a given threshold. (𝑉, 𝐸) be an undirected graph with a vertex set 𝑉 and an
4 Mathematical Problems in Engineering

z2 where 𝑍 is the normalization constant, 𝜙𝑖 (𝑥𝑖 , 𝑧𝑖 ) is the node

potential at the node 𝑖, and 𝜓𝑖𝑗 (𝑥𝑖 , 𝑥𝑗 ) is the edge potential
x3
2 (x2 , z2 ) between the nodes 𝑖 and 𝑗.
)
12 (x1 , x2 ) 2
(x 2
m 3→ 3.3. Belief Propagation. The task of inference is to determine
x1 x2 N (x2 ) \ x1
the underlying state of the random variables in the graphical
m2→1 (x1 )
m
model. A key problem in designing algorithms for inference
4→ x4 is finding an efficient way to reason about the large number
2 (x
2)
of possible assignments to the variables in the model. One
Figure 3: A graph fragment showing the updating rule of the fundamental inference problem is to find the most likely
message sent from node 2 to node 1. configuration of the probability distribution, which is known
as the MAP assignment. Many NP-hard combinatorial opti-
mization problems (e.g., MWISP) can be posed as finding
edge set 𝐸. The edge (𝑖, 𝑗) is included in the graph only if two the MAP inference problem in pairwise Markov random
nodes 𝑖 and 𝑗 are connected to each other. The independent fields with binary variables. There are many ways to solve
set on the graph is defined as a subset of 𝑉 that has no edge the MAP inference problem. The belief propagation (BP)
between any two nodes. A binary vector x, the length of which algorithm, specifically the max-product belief propagation
equals the cardinality of 𝑉, that is, |𝑉|, is defined to represent (MPBP) algorithm, is an efficient way to solve MAP inference
an independent set of the graph 𝐺. 𝑥𝑖 takes value 1 if node 𝑖 problems in graphical models.
belongs to the independent set and 0 otherwise. Define 𝑤𝑖 as MPBP is an approximate inference algorithm that is
the weight on node 𝑖; then MWISP is to find an independent simple to code and scales very well with problem size. By
set of 𝑉 that the sum of corresponding weights is largest; that taking advantage of the statistical independence in the graph
is, structure, MPBP algorithm can break the global inference
problem into localized operations, which is generally faster
max
𝑥
∑ 𝑤𝑖 𝑥𝑖 , than manipulating the joint distribution explicitly. The MPBP
𝑖∈𝑉 algorithm operates by a message passing mechanism which
(5)
s.t. 𝑥𝑖 + 𝑥𝑗 ≤ 1, ∀ (𝑖, 𝑗) ∈ 𝐸. is implemented by iteratively passing messages along edges of
the graph that summarize each variable’s beliefs. After receiv-
ing messages from its neighbors, the node updates its belief
3.2. Graphical Models. The core component of graphical and propagates it to the rest of the graph. An assignment can
models is the representation of a probability distribution then be decided from the beliefs by choosing the most likely
using a graph as a data structure. A graphical model is state according to each node’s belief. The MPBP algorithm
denoted as a graph 𝐺 = (𝑉, 𝐸) with a probability distribution makes direct use of the graph structure in constructing and
𝑝(x), in which nodes represent the variables x = {𝑥𝑖 | 𝑖 ∈ passing messages, making it simple to implement and run
𝑉}, and edges represent the probabilistic interaction between quickly. As a result, the review will primarily focus on the
the neighboring variables. Basically there are two types of MPBP algorithm, which approximates the MAP inference
graphical representation of distributions. One is the Bayesian problem, as it will be used in the proposed approach.
network which utilizes a directed graph; and the other is the The update equations for MPBP algorithm can be derived
Markov Random Field (MRF) using an undirected graph. from the probability distribution. The messages are updated
As the graph model used in the MPBP-MHT approach is as follows: each node sends messages to its neighbors and
the undirected graph, in this paper we focus on the MRF. A receives messages from them and the messages are updated
graph is called an MRF if every variable 𝑥𝑖 is independent according to the received messages from neighbors at the
of nonneighboring variables in the graph given the value previous iteration
of its neighbors. For instance, in Figure 3, the lack of an
edge between nodes 𝑥1 and 𝑥3 means they are conditionally 𝑚𝑖→𝑗 (𝑥𝑗 )
independent under a given 𝑥2 .
(7)
A clique of a graph is a subset of this graph in which all = max 𝜓𝑖𝑗 (𝑥𝑖 , 𝑥𝑗 ) 𝜙𝑖 (𝑥𝑖 , 𝑧𝑖 ) ∏ 𝑚𝑘→𝑖 (𝑥𝑖 ) ,
𝑥𝑖
the nodes are fully connected. A pairwise Markov network 𝑥𝑘 ∈𝑁(𝑥𝑖 )\𝑥𝑗
is an undirected graphical model that only takes the edges
as cliques, which means that the potentials are over single where 𝑚𝑖→𝑗 (𝑥𝑗 ) represents the message passed from node 𝑥𝑖
variables or pairs of variables. More precisely, let x denote to node 𝑥𝑗 and 𝑁(𝑥𝑖 ) is the set of all the neighbors of node 𝑥𝑖 .
the values of all unobserved variables in the graph and we At any iteration, once a node has received all messages
assume that each unobserved node 𝑥𝑖 is associated with an from its neighbor nodes, the belief at node 𝑥𝑖 can be
observation 𝑧𝑖 . Then the probability distribution 𝑝(x | z) computed by
of a pairwise MRF is expressed in terms of the product of
potentials on edges and nodes 𝑏𝑖 (𝑥𝑖 ) = 𝜙𝑖 (𝑥𝑖 , 𝑧𝑖 ) ∏ 𝑚𝑘→𝑖 (𝑥𝑖 ) , (8)
𝑥𝑘 ∈𝑁(𝑥𝑖 )
1
𝑝 (x | z) = ∏𝜓 (𝑥 , 𝑥 ) ∏𝜙 (𝑥 , 𝑧 ) , (6) and the output of (8) should be normalized. Initially, the mes-
𝑍 𝑖,𝑗 𝑖𝑗 𝑖 𝑗 𝑖 𝑖 𝑖 𝑖
sages are initialized with constant functions. As the observed
Mathematical Problems in Engineering 5

nodes are localized and do not receive messages, generally we

1 Track 1
denote 𝜙𝑖 (𝑥𝑖 , 𝑧𝑖 ) as 𝜙𝑖 (𝑥𝑖 ) for notation simplicity.

4. Efficient Multitarget Tracking Using

Graphical Models
4 Track 4 3 Track 3 2 Track 2
The main reason for constructing graphical model repre-
sentation for data association is to utilize efficient inference
algorithms, specifically, the max-product belief propagation
algorithm, to seek the best global hypothesis. In this section,
we first present the approach we take to construct graphical 5 Track 5
models for data association problem in MTT applications.
Then we describe the detailed formulation of the proposed Figure 4: Graphical model for data association problem in Figure 2.
MPBP-MHT method. Specific process of defining potential
functions and running message passing algorithms is also
given. track score and in our approach we only use the tracks that
have positive track scores. An edge is added between nodes
4.1. Graphical Model Representation. TOMHT seeks and gen- if the corresponding tracks are incompatible. Therefore, the
erates the best global hypothesis by enumerating all possible process of seeking the best global hypothesis can be translated
associations over time which, as we can see, is equal to to finding the maximum weight independent set on the
seeking maximum a posteriori (MAP) assignment over the graph.
report data. Let T and z denote the set of all tracks and Taking the track trees in Figure 2 as an example, it is
observations up to the present scan, respectively. A binary clear that 5 track hypotheses are formulated up to scan 𝑘.
vector h, the length of which equals the cardinality of T, All tracks that share one or more common observations are
that is, |T|, is used to denote the global hypothesis where connected to each other with edges in the corresponding
ℎ𝑖 = 1 if the 𝑖th track is included in the global hypothesis and graph in Figure 4. Assuming the track score of track 1 to track
ℎ𝑖 = 0 otherwise. In this way, the process of finding the best 5 is one to five, respectively (the score is labeled within the
hypothesis h∗ is equivalent to computing the MAP solution cycles), in this case the best global hypothesis with the largest
of the posterior probability distribution 𝑝(h | z) sum of weights consists of tracks 2, 3, and 4.
h∗ = arg max 𝑝 (h | z) . (9) 4.2. MPBP-MHT Algorithm for Data Associations. Once a
h
graphical model is constructed, the specific form of the
The specific logarithmic formula of 𝑝(h | z) is given by Kurien potential functions and the update equations for the max-
in [13] product belief propagation algorithm can be calculated. As
the joint probability distribution is consistent with a pairwise
|T| MRF, it can be factored into a product of potential functions
log 𝑝 (h | z) = 𝐶 + ∑ℎ𝑖 𝑠𝑖 , (10) and thus the probability distribution 𝑝(h | z) of hypothesis
𝑖=1
variable h given an observation set z is
where 𝐶 is a constant and 𝑠𝑖 is the track score of the 𝑖th track. 1
Therefore, the process of computing the MAP solution of 𝑝 (h | z) = ∏ 𝜓 (ℎ , ℎ ) ∏𝜙 (ℎ ) , (12)
𝑝(h | z) can be formulated as an integer linear program which 𝑍 (𝑖,𝑗)∈𝐸 𝑖𝑗 𝑖 𝑗 𝑖 𝑖 𝑖
is subject to the restriction that no common observation is
shared by tracks included in the same global hypothesis where the edge potentials 𝜓𝑖𝑗 (ℎ𝑖 , ℎ𝑗 ) between nodes 𝑖 and 𝑗 are

|T|
{0 if ℎ𝑖 = 1, ℎ𝑗 = 1
h∗ = arg max h ⋅ s = arg max∑ℎ𝑖 𝑠𝑖 . (11) 𝜓𝑖𝑗 (ℎ𝑖 , ℎ𝑗 ) = { (13)
h h
𝑖=1 1 otherwise.
{
Due to the equivalence of the integer linear program Assume 𝑠𝑖 is the track score of the 𝑖th track; then the potential
problem in (11) and MWISP described in (5), the graph function 𝜙𝑖 (ℎ𝑖 ) at node 𝑖 is
representation for data association can be constructed by
utilizing the graphical model in MWISP so that efficient 𝑠
inference algorithms can be used to calculate the MAP {𝑒 𝑖 if ℎ𝑖 = 1
𝜙𝑖 (ℎ𝑖 ) = { (14)
assignment in MTT application. 1 otherwise.
{
In individual clusters, we construct the graph 𝐺𝑘 =
(𝑉 , 𝐸𝑘 ) at scan 𝑘. Every node represents a possible track and
𝑘
It is clear that if h is an independent set, then 𝑝(h | z) =
is associated with a hypothesis variable ℎ𝑖 . The variable ℎ𝑖 = 1 (1/𝑍) exp(∑𝑖 ℎ𝑖 𝑠𝑖 ); otherwise, 𝑝(h | z) equals zero. If h∗ =
if the track is included in the global hypothesis and ℎ𝑖 = 0 arg maxh 𝑝(h | z), then the MAP assignment corresponds
corresponds to absence. The weight of a node is defined as the to a maximum weight independent set in the graph. It is
6 Mathematical Problems in Engineering

clear that finding the MWIS is equivalent to finding the MAP until h𝑛 converges or the number of iterations exceeds a given
assignment on the corresponding graphical model. Therefore, threshold.
the max-product belief propagation algorithm can be used The specific pseudocode of the MPBP-MHT algorithm
as an iterative strategy for finding the MWIS solution on the is shown in Algorithm 1. The hypothesis generation step is
graph. executed in independent cluster. Input parameters include
The main idea of max-product belief propagation is to the current graph 𝐺 and scan index 𝑘 is omitted for simplicity.
convert messages between nodes iteratively, which aims to In the practical tracking application, the graphical models
maximize the joint probability by finding the most likely corresponding to the track trees can be singly connected
assignment to all of the variables. In each iteration, each graphs (i.e., there is only one path between any two given
node sends messages to its neighbors and the belief can be nodes and in this case the graph structure resembles a
updated by exploiting the incoming messages at each node. tree) or graphs with cycles. For a tree-structured graph, the
This procedure is repeated until convergence. max-product belief propagation algorithm is guaranteed to
𝑛 𝑛 𝑛
Define 𝑚𝑖→𝑗 = [𝑚𝑖→𝑗 (0), 𝑚𝑖→𝑗 (1)]𝑇 as the messages converge in a finite number of iterations and the assignment
passed from the node 𝑖 to its neighbor node 𝑗 in the 𝑛th itera- based on the messages at convergence is guaranteed to give
0 the optimal assignment values corresponding to the MAP
tion. Initialize the iteration by setting 𝑛 = 0 and 𝑚𝑖→𝑗 =
0 0 𝑇 solution [14]. As for a graph with cycles, its performance has
[𝑚𝑖→𝑗 (0), 𝑚𝑖→𝑗 (1)] , where not been well studied and the algorithm may not converge.
Nevertheless, max-product algorithms often show remark-
0 {exp (𝑠𝑖 ) if 𝑟 = 0 able performances even on graphs with cycles. Recently
𝑚𝑖→𝑗 (𝑟) = { (15) several good empirical performances have been obtained by
1 otherwise.
{ running the max-product algorithm on loopy graphs [15–18]
and progress has been made to understand the mechanism
For 𝑛 ≥ 1, in each iteration, the message in a node is updated
behind its empirical success. For loopy graph with a single
according to the messages from its neighbors recursively,
loop, [19] demonstrated that the algorithm converges to the
where the updating rule is
correct marginal or MAP probabilities. For arbitrary graphs,
𝑛 𝑛−1 [20] proved that the assignment based on a fixed point is a
𝑚𝑖→𝑗 = max 𝜓𝑖𝑗 (ℎ𝑖 , ℎ𝑗 ) 𝜙𝑖 (ℎ𝑖 ) ∏ 𝑚𝑘→𝑖 . (16)
ℎ𝑖
𝑘∈𝑁(𝑖)\𝑗
neighborhood maximum of the posterior probability.

Specifically, by combining (13), (14), and (16), the updating 5. Experimental Results
rule can be rewritten as
𝑛
In this section, we conduct an empirical evaluation of the
𝑚𝑖→𝑗 (0) proposed MBPB-MHT algorithm over a challenging situa-
tion using simulated data. We present the setup and scenarios
under which we operate our experiments. The performance
= max [ ∏ 𝑚𝑘→𝑖 𝑛−1
(0) , exp (𝑠𝑖 ) ∏ 𝑚𝑘→𝑖 𝑛−1
(1)] , of the proposed approach was evaluated with several per-
(17)
𝑘∈𝑁(𝑖)\𝑗 𝑘∈𝑁(𝑖)\𝑗
[ ] formance metrics. To further demonstrate its superiority,
𝑛 𝑛−1 comparisons were made with GRASP-MHT algorithm in
𝑚𝑖→𝑗 (1) = ∏ 𝑚𝑘→𝑖 (0) . correlation quality, overall cardinality, and state estimation.
𝑘∈𝑁(𝑖)\𝑗

Each node 𝑖 maintains a belief 𝑏𝑖𝑛 = [𝑏𝑖𝑛 (0), 𝑏𝑖𝑛 (1)], which is 5.1. Target Motion and Measurement Models. The accelera-
computed as follows: tion model with a white Gaussian noise is considered and
the linear Kalman filter is used in our experiments. Targets
𝑏𝑖𝑛 = 𝜙𝑖 (ℎ𝑖 ) ∏ 𝑚𝑘→𝑖
𝑛
, (18)
move in a 2D surveillance area. The true state of a target at
𝑇
𝑘∈𝑁(𝑖) ̇
scan 𝑘 is x𝑘 = [𝑥(𝑘), 𝑦(𝑘), 𝑥(𝑘), ̇
𝑦(𝑘)] . 𝑥(𝑡𝑘 ) and 𝑦(𝑡𝑘 ) are,
respectively, the positions of this target in the 𝑋-𝑌 coordinate
where in the same way the specific form is system, while 𝑥(𝑡 ̇ 𝑘 ) are, respectively, the velocities
̇ 𝑘 ) and 𝑦(𝑡
of this target. The movement of each target is modeled as the
𝑏𝑖𝑛 (0) = ∏ 𝑚𝑘→𝑖
𝑛
(0) , following target motion model:
𝑘∈𝑁(𝑖)
(19)
x𝑘 = Fx𝑘−1 + k𝑘 , (21)
𝑏𝑖𝑛 (1) = exp (𝑠𝑖 ) ∏ 𝑚𝑘→𝑖
𝑛
(1) .
𝑘∈𝑁(𝑖)
where the state transition matrix F is expressed as
At the end of the 𝑛th iteration, the maximum weight inde-
pendent set h𝑛 is estimated as 1 0 𝑇 0
[ ]
ℎ𝑖𝑛 = 1{𝑏𝑖𝑛 (1)>𝑏𝑖𝑛 (0)} . [0 1 0 𝑇]
(20) [
F=[ ], (22)
]
[0 0 1 0 ]
The track hypothesis will be included in the MWIS if its
belief satisfies 𝑏𝑖𝑛 (1) > 𝑏𝑖𝑛 (0). The iteration process will repeat [0 0 0 1 ]
Mathematical Problems in Engineering 7

Input: graph 𝐺, including a cluster of family, track score and track ICL.
Output: best global hypothesis
0 0 0
(1) initialization: set 𝑚𝑖→𝑗 = [𝑚𝑖→𝑗 (0), 𝑚𝑖→𝑗 (1)]𝑇 with (15).
(2) iteration: At iteration n for all nodes 𝑖 ∈ 𝑉
𝑛
(a) Calculate new message 𝑚𝑖→𝑗 which is sent by the node 𝑖 to all its neighbors with (17).
(b) Calculate the belief at each node with (19).
(c) Decision: for each node 𝑖 ∈ 𝑉, compare 𝑏𝑖𝑛 (1) and 𝑏𝑖𝑛 (0); if 𝑏𝑖𝑛 (1) > 𝑏𝑖𝑛 (0), set ℎ𝑖𝑛 = 1.
(3) if h𝑛 converges, finish the iteration and output h∗ ; else set 𝑛 = 𝑛 + 1 and go to step (2)

Algorithm 1: MPBP-MHT algorithm. Note. For computational stability, it is often recommended that messages should be normalized in
each iteration.

and the sampling period T = 2 s. The covariance matrix of the standard deviation of the process noise is 10 m. The target
white Gaussian noise k𝑘 is detection probability is 0.9 and the false alarm density is set
as 10−8 /m2 .
𝑇3 𝑇2
[3 0 2 0]
[ ] 5.2.2. Scenario B. The scenario B consists of ten closely
[ 3
𝑇2 ]
[0 𝑇 0 ] spaced targets and the targets are moving in formation with
[ 3 2 ]
[
Q=[ 2 ] 𝑞, (23) a separation of 900 m and a speed of 300 m/s over a period
]
[𝑇 0 𝑇 0]
of 120 s. The start and stop positions are marked with ∘ and
[ ] ×, respectively. Each target performs three turns with an
[2 ]
[ 𝑇 2 ] acceleration of 3𝑔 at 22 s, 42 s, and 72 s, where 𝑔 is the gravita-
0 0 𝑇
[ 2 ] tional acceleration and each turn lasts for 8 s with a course
change of 45o . The target detection probability is set as 0.9
where the process noise intensity 𝑞 = 100 m2 /s3 . The meas- and the false alarm density is 10−8 /m2 . Figure 6(a) shows the
urement model is given as true target trajectories in the scenario B used for performance
evaluation and Figure 6(b) provides the true observations
z𝑘 = Hx𝑘 + w𝑘 , (24)
with clutter. The number of clutter is Poisson distributed
where and their locations in the observation space are uniformly
distributed.
1 0 0 0
H=[ ], (25) 5.3. Results and Evaluation. In this part, we provide the
0 1 0 0
empirical results in the simulation. In order to obtain a quan-
and the zero-mean Gaussian measurement noise w𝑘 has a titative assessment of the performance of two algorithms, a set
covariance matrix R = diag{𝜎𝑥2 , 𝜎𝑦2 }, where 𝜎𝑥2 = 𝜎𝑦2 = 50 m2 . of performance metrics, including miscorrelation rate, cor-
rect correlation rate, average time for hypothesis per scan, and
5.2. Simulation Scenarios. We employ two scenarios for the optimal subpattern assignment distance, are introduced
verifying the performance of the proposed method. All the and the detailed description could be found in [21].
estimation results were based on 100 Monte Carlo runs and
the depth of TOMHT equals 5. The tracking performance (a) Miscorrelation rate of true tracks (𝑅MC ): we use the
is then compared with the GRASP-MHT algorithm in the miscorrelation rate to measure the data association
same scenario, respectively. In the GRASP-MHT algorithm, quality. The miscorrelation rate is defined as the ratio
the maximum number of tuples is set to 30 and the number of average number of miscorrelation over the average
of randomized iterations for each tuple is set to 3. All other track life.
parameters were shared with MPBP-MHT algorithm. (b) Correct correlation rate of true tracks (𝑅CC ): the
correct correlation rate is also used for evaluating
5.2.1. Scenario A. In the scenario A, the number of scans is set
the correctness of data association. It is defined as
to 80. At the beginning, two closely spaced targets are moving
the ratio of total number of correctly associated
parallelly with a separation of 30 m and a speed of 30 m/s over
observations in true tracks to the total number of
a period of 30 s. In the later 50 s, there is an intersect at 32 s
target-originated observations.
and after that two targets are moving separately. Figure 5(a)
shows the real tracks of the two targets, and Figure 5(b) shows (c) Average time for hypothesis per scan (𝑇𝐻): this metric
the real observations with clutters. The observation errors of is defined to evaluate the computational complexity of
azimuth and range are 0.002 rad and 20 m, respectively. The different trackers with unity being second.
8 Mathematical Problems in Engineering

2000 2000

1500 1500

1000 1000

500 500
y (m)

y (m)
0 0

−500 −500

−1000 −1000

−1500 −1500
−500 0 500 1000 1500 2000 2500 3000 −500 0 500 1000 1500 2000 2500 3000
x (m) x (m)
(a) (b)

Figure 5: Target trajectories and the real observations with clutters in scenario A. (a) Target trajectories; (b) real observations.

40000 40000

35000 35000

30000 30000

25000 25000
y (m)
y (m)

20000 20000

15000 15000

10000 10000

5000 5000

0 0
−25000 −20000 −15000 −10000 −5000 0 5000 −25000 −20000 −15000 −10000 −5000 0 5000
x (m) x (m)
(a) (b)

Figure 6: Target trajectories and the real observations with clutters in scenario B. (a) Target trajectories; (b) real observations.

(d) The optimal subpattern assignment (OSPA) distance: Table 1: Simulation results of the scenario A.
this metric is introduced to measure the quality of
cardinality and state estimation. The OSPA distance Performance metrics
Tracker
between the two sets 𝑋 = {𝑥1 , . . . , 𝑥𝑛 } and 𝑌 = {𝑦1 , 𝑅MC 𝑅CC 𝑇𝐸 (𝑠)
. . . , 𝑦𝑚 } is calculated by MPBP-MHT 0.01 0.85 0.002
(𝑐)
𝑑ospa (𝑋, 𝑌) GRASP-MHT 0.24 0.67 0.001
𝑝
{ 1
{( ( min max 𝑑(𝑐) (𝑥𝑖 , 𝑦̂𝜋(𝑖) )𝑝 + 𝑐𝑝 (𝑛 − 𝑚))) 𝑛 ≥ 𝑚 (26)
= { 𝑛 𝜋∈∏ 𝑛 1≤𝑖≤𝑛
{ (𝑐)
𝑑 (𝑌, 𝑋) 𝑛 ≤ 𝑚, trajectories by GRASP-MHT and have less track confusion.
{ ospa
Additionally, from Table 1, it can be seen that the correct
where ∏ 𝑛 denotes the set of all possible permutations correlation rate 𝑅CC of MPBP-MHT is larger than that of
of {1, 2, . . . , 𝑛} and 𝑑(𝑐) (𝑥, 𝑦) = min(𝑐, 𝑑(𝑥, 𝑦)) is the GRASP-MHT and the miscorrelation rate 𝑅MC of MPBP-
truncated Euclidean distance between the vectors 𝑥 MHT is smaller than that of GRASP-MHT, which means
and 𝑦. The cut-off distance 𝑐 is set as 5000 and the that the correctness of data association in MPBP-MHT is
order parameter 𝑝 is fixed at 2. better than GRASP-MHT. The quantitative difference of
𝑇𝐻 between the two algorithms is merely 0.001 s, which
5.3.1. Scenario A. The performance of the two algorithms indicates that both algorithms can operate efficiently. Finally,
in scenario A is shown in Table 1 and Figures 7 and 8. At the overall tracking performance is compared in terms of
first glance, we can observe from Figure 7 that the estimated the OSPA distance. The OSPA distance is used to evaluate
trajectories produced by MPBP-MHT are smoother than both cardinality and state estimation. It can be seen that
Mathematical Problems in Engineering 9

2000 2000

1500 1500

1000 1000

500 500

y (m)
y (m)

0 0

−500 −500

−1000 −1000

−1500 −1500
−500 0 500 1000 1500 2000 2500 3000 −500 0 500 1000 1500 2000 2500 3000
x (m) x (m)
(a) (b)

Figure 7: Estimated trajectories in scenario A. (a) MPBP-MHT; (b) GRASP-MHT.

2500

2000
OSPA distance

1500

1000

500

0
0 10 20 30 40 50 60 70 80
Frame number

MPBP-MHT
GRASP-MHT

Figure 8: OSPA distance of the scene A.

Table 2: Simulation results of the scene B. tracker using the MPBP-MHT algorithm gets cleaner tracks
while the tracker using the GRASP-MHT algorithm exhibits
Performance metrics obvious track switching.
Tracker
𝑅MC 𝑅CC 𝑇𝐸 (𝑠) Table 2 summarizes the tracking performance of the two
MPBP-MHT 0.02 0.78 0.16 algorithms in correlation quality and execution time. From
the statistics, the MPBP-MHT algorithm obtains a larger
GRASP-MHT 0.17 0.69 0.04
𝑅CC and a smaller 𝑅MC , which demonstrates that the MPBP-
MHT algorithm outperforms the GRASP-MHT algorithm
in the data association part. On timing results, both of the
although the OSPA distance of MPBP-MHT in some frames algorithms are efficient while the GRASP-MHT algorithm
is larger than that of GRASP-MHT, in general, the tracking has a comparatively shorter execution time.
performance of MPBP-MHT outperforms GRASP-MHT in Figure 10 shows the average OSPA distance obtained
terms of the OSPA distance with great priority. over 100 Monte Carlo runs. It can be seen that the OSPA
distance of GRASP-MHT fluctuates approximately between
5.3.2. Scenario B. The performance of the two algorithms 1250 and 3000 whereas the OSPA distance of MPBP-MHT
in scenario B is shown in Table 2 and Figures 9 and 10. fluctuates between 250 and 2600. Although the curves of the
Figure 9 shows the estimated tracks obtained by MPBP-MHT two algorithms follow a similar pattern, the curve of MPBP-
and GRASP-MHT in scenario B, respectively. It can be seen MHT is beneath the curve of GRASP-MHT most of time and
that MPBP-MHT does better in tracking performance. The stabilizes at a lower position. Therefore, we can conclude that
10 Mathematical Problems in Engineering

40000 40000

35000 35000

30000 30000

25000 25000
y (m)

y (m)
20000 20000

15000 15000

10000 10000

5000 5000

0 0
−25000 −20000 −15000 −10000 −5000 0 5000 −25000 −20000 −15000 −10000 −5000 0 5000
x (m) x (m)
(a) (b)

Figure 9: Estimated trajectories in scenario B. (a) MPBP-MHT; (b) GRASP-MHT.

3500 Acknowledgments
3000 This work was supported by the National Natural Science
2500
Foundation of China (Grant no. 61471019).
OSPA distance

2000 References
1500 [1] D. B. Reid, “An algorithm for tracking multiple target,” IEEE
Transactions on Automatic Control, vol. 24, no. 6, pp. 843–854,
1000 1979.
500 [2] A. Frank, P. Smyth, and A. Ihler, “Beyond MAP estimation with
the track-oriented multiple hypothesis tracker,” IEEE Trans-
0 actions on Signal Processing, vol. 62, no. 9, pp. 2413–2423, 2014.
0 50 100 150
[3] A. B. Poore and N. Rijavec, “Multitarget tracking and multidi-
Frame number
mensional assignment problems,” in Proceedings of the Signal
MPBP-MHT and Data Processing of Small Targets, pp. 345–356, April 1991.
GRASP-MHT [4] C.-Y. Chong, “Graph approaches for data association,” in Pro-
ceedings of the 15th International Conference on Information
Figure 10: OSPA distance of the scene B. Fusion, FUSION 2012, pp. 1578–1585, September 2012.
[5] H. Wang, J. Sun, S. Lu, and S. Wei, “Factor graph aided multiple
hypothesis tracking,” Science China Information Sciences, vol.
the MPBP-MHT algorithm exhibits better overall estimation 56, no. 10, pp. 1876–1887, 2013.
performance. [6] D. J. Papageorgiou and M. R. Salpukas, The Maximum Weight
Independent Set Problem for Data Association in Multiple Hypo-
6. Conclusions thesis Tracking, vol. 381 of Lecture Notes in Control and Informa-
tion Sciences, Springer, Berlin, Germany, 2009.
In this paper, we propose an efficient MPBP-MHT method [7] X. Ren, Z. Huang, S. Sun, D. Liu, and J. Wu, “An efficient MHT
that exploits graphical models and message passing algo- implementation using GRASP,” IEEE Transactions on Aerospace
rithms to solve the data association problem in tracking appli- and Electronic Systems, vol. 50, no. 1, pp. 86–101, 2014.
cation. Instead of seeking the best hypothesis by enumeration, [8] A. Yuille, “Belief Propagation, Mean Field, and Bethe Approx-
we cast the hypothesis generation problem into the graphical imations,” in In Advances in Markov Random Fields for Vision
model formalism using MWISP structure and the efficient and Image Processing, MIT Press, Cambridge, Mass, USA, 2011.
MPBP algorithm is developed to find the MAP assignment. [9] Y. Bar-Shalom, S. S. Blackman, and R. J. Fitzgerald, “Dimen-
The effectiveness of the proposed algorithm is tested over sionless score function for multiple hypothesis tracking,” IEEE
challenging cases and a comparison with the GRASP-MHT Transactions on Aerospace and Electronic Systems, vol. 43, no. 1,
algorithm is made to demonstrate the priority in association pp. 392–400, 2007.
property of the MPBP-MHT algorithm. [10] S. S. Blackman and R. Popoli, Design and Analysis of Modern
Tracking Systems, Artech House, Norwood, MA, USA, 1999.
Conflicts of Interest [11] J. Fu, J. Sun, S. Lu, and Y. Zhang, “Multiple hypothesis tracking
based on the Shiryayev sequential probability ratio test,” Science
The authors declare that they have no conflicts of interest. China Information Sciences, vol. 59, no. 12, pp. 1–11, 2016.
Mathematical Problems in Engineering 11

[12] T. Sathyan, T.-J. Chin, S. Arulampalam, and D. Suter, “A mul-

tiple hypothesis tracker for multi-target tracking with multiple
simultaneous measurements,” IEEE Journal on Selected Topics in
Signal Processing, vol. 7, no. 3, pp. 448–460, 2013.
[13] T. Kurien, “Issues in the Design of Practical Multitarget-Multi-
sensor Tracking: Advanced and Applications,” in Chapter 3,
Artech House, Boston, MA, USA, 1990.
[14] M. Bayati, D. Shah, and M. Sharma, “Maximum weight match-
ing via max-product belief propagation,” in Proceedings of the
2005 IEEE International Symposium on Information Theory, ISIT
05, pp. 1763–1767, September 2005.
[15] K. Murphy, Y. Weiss, and M. I. Jordan, “Loopy belief propa-
gation for approximate inference: an empirical study,” in Pro-
ceedings of the Fifteenth Conference on Uncertainty in Artifi-
cial Intelligence, pp. 467–475, Morgan Kaufmann Publishers,
Burlington, Mass, USA, 1999.
[16] L. Chen, M. J. Wainwright, M. Çetin, and A. S. Willsky, “Data
association based on optimization in graphical models with
application to sensor networks,” Mathematical and Computer
Modelling, vol. 43, no. 9-10, pp. 1114–1135, 2006.
[17] W. T. Freeman, E. C. Pasztor, and O. T. Carmichael, “Learning
low-level vision,” International Journal of Computer Vision, vol.
40, no. 1, pp. 25–47, 2000.
[18] C.-H. Huang, Y. Li, and L. Dolecek, “Belief propagation algo-
rithms on noisy hardware,” IEEE Transactions on Communica-
tions, vol. 63, no. 1, pp. 11–24, 2015.
[19] Y. Weiss and W. T. Freeman, “Correctness of belief propagation
in gaussian graphical models of arbitrary topology,” Neural
Computation, vol. 13, no. 10, pp. 2173–2200, 2001.
[20] Y. Weiss and W. T. Freeman, “On the optimality of solutions
of the max-product belief-propagation algorithm in arbitrary
graphs,” IEEE Transactions on Information Theory, vol. 47, no. 2,
pp. 736–744, 2001.
[21] H. Leung, H. U. Zhijian, and M. Blanchette, “Evaluation of
multiple radar target trackers in stressful environments,” IEEE
Transactions on Aerospace and Electronic Systems, vol. 35, no. 2,
pp. 663–674, 1999.
Advances in Advances in Journal of Journal of
Operations Research
Hindawi Publishing Corporation
Decision Sciences
Hindawi Publishing Corporation
Applied Mathematics
Hindawi Publishing Corporation
Algebra
Hindawi Publishing Corporation
Probability and Statistics
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

The Scientific International Journal of

World Journal
Hindawi Publishing Corporation
Differential Equations
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Submit your manuscripts at

https://www.hindawi.com

International Journal of Advances in

Combinatorics
Hindawi Publishing Corporation
Mathematical Physics
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of Journal of Mathematical Problems Abstract and Discrete Dynamics in

Complex Analysis
Hindawi Publishing Corporation
Mathematics
Hindawi Publishing Corporation
in Engineering
Hindawi Publishing Corporation
Applied Analysis
Hindawi Publishing Corporation
Nature and Society
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

International
Journal of Journal of
Mathematics and
Mathematical
#HRBQDSDĮ,@SGDL@SHBR
Sciences

Journal of International Journal of Journal of

Hindawi Publishing Corporation Hindawi Publishing Corporation Volume 2014

Function Spaces
Hindawi Publishing Corporation
Stochastic Analysis
Hindawi Publishing Corporation
Optimization
Hindawi Publishing Corporation
http://www.hindawi.com Volume 201 http://www.hindawi.com http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014